Running the test-suite terminates abruptly with fd:5: hGetLine: end of file in MSYS2 Windows

[jasagredo]

Describe the bug
The test-suite terminates non-deterministically with:

cabal-tests.exe: fd:5: hGetLine: end of file

To Reproduce
Steps to reproduce the behavior:

$ ./validate.sh -v -s build -s lib-suite
...
=== Cabal: cabal-testsuite =================================================== 00:10:15 ===
C:\Users\Javier\code\cabal\dist-newstyle-validate-ghc-9.6.3\build\x86_64-windows\ghc-9.6.3\cabal-testsuite-3\build\cabal-tests\cabal-tests.exe --builddir=/c/Users/Javier/code/cabal/dist-newstyle-validate-ghc-9.6.3/build/x86_64-windows/ghc-9.6.3/cabal-testsuite-3 -j4 --with-ghc=ghc
threads: 4
tests to run: 509
PackageTests\TestSuiteTests\ExeV10\cabal-with-hpc.multitest.hs                                                    SKIP no cabal-install (4.38s)
PackageTests\AllowOlder\cabal.test.hs                                                                             SKIP no cabal-install (4.45s)
PackageTests\AllowNewer\cabal.test.hs                                                                             SKIP no cabal-install (4.45s)
PackageTests\AutoconfBadPaths\cabal.test.hs                                                                       SKIP no cabal-install (0.53s)
PackageTests\AutogenModules\Package\setup.test.hs                                                                 OK (1.51s)
PackageTests\AutogenModules\SrcDist\setup.test.hs                                                                 OK (2.23s)
PackageTests\Backpack\bkpcabal01\cabal.test.hs                                                                    SKIP no cabal-install (0.54s)
...
SKIP no cabal-install
END Test.Cabal.Server
GHCi exited with ExitFailure (-1073741510) (use -v for more information)
cabal-tests.exe: fd:4: hGetLine: end of file

Turns out that annotating the readUntilEnd and readUntilSigil calls to hGetLine shows that non-deterministically, one of them fails. If I also print the script_path provided, it is different each time, even sometimes it refers to tests which are marked as skipIfWindows.

I don't know how the Test.Cabal.Server is setup, maybe we are leaving the GHCi session in a zombie state?

Expected behavior
The test-suite should just run normally, as it does on CI.

System information

• MINGW64_NT-10.0-22621 3.4.10.x86_64 Msys
• cabal 3.10.2.0, ghc 9.6.3

[jasagredo]

I was suggested to use the RTS flag --io-manager=native. This fixes the problem 🎉

I will leave this issue open to make sure that ./validate.sh takes this into account in some way.

[jasagredo]

(Thanks for the GHC team at IOG for the suggestion @hsyl20 😄)

[ulysses4ever]

Still happening on master, unfortunately.

[jasagredo]

This is starting to be beyond my knowledge. I re-enabled the Windows CI in #10282 with the following results:

• First a cabal-tests.exe: fd:5: hGetLine: end of file on 9.10.1
  logs-failed-9.10.1.txt

• In the next run, cabal-tests.exe: fd:5: hGetLine: end of file on 9.8.2
  logs-failed-9.8.2.txt and cabal-tests.exe: fd:8: hGetLine: end of file in 9.6.4
  logs-failed-9.6.4.txt

• Then I enabled --io-manager=native which solved it for my local runs, resulting in a segfault in 9.0.2:

  2024-08-27T23:04:42.4099278Z 
  2024-08-27T23:04:42.4099909Z Access violation in generated code when writing 0x0
  2024-08-27T23:04:42.5100373Z 
  2024-08-27T23:04:42.5100863Z  Attempting to reconstruct a stack trace...
  2024-08-27T23:04:42.5101196Z 
  2024-08-27T23:04:42.5101349Z    Frame	Code address
  2024-08-27T23:04:42.5102665Z  * 0x9a215dd20	0x28ae318 D:\a\cabal\cabal\dist-newstyle-validate-ghc-9.0.2\build\x86_64-windows\ghc-9.0.2\cabal-testsuite-3\build\cabal-tests\cabal-tests.exe+0x24ae318
  2024-08-27T23:04:42.5108006Z  * 0x9a215dd28	0x2769608 D:\a\cabal\cabal\dist-newstyle-validate-ghc-9.0.2\build\x86_64-windows\ghc-9.0.2\cabal-testsuite-3\build\cabal-tests\cabal-tests.exe+0x2369608
  2024-08-27T23:04:42.5109223Z  * 0x9a215dd30	0x345e680
  2024-08-27T23:04:42.5109543Z  * 0x9a215dd38	0x7ef4fe1057b0
  2024-08-27T23:04:42.5109871Z  * 0x9a215dd40	0x7ef531e058d8
  

  logs-segfault-9.0.2.txt

• Then I enabled --io-manager=native for all but 9.0.2, resulting in a cabal-tests.exe: fd:8: hGetLine: end of file exception in 9.0.2 logs-segfault-another-9.0.2.txt

Notice it is not the same test that throws this exception. I'm running out of ideas here so I gently invoke @Mistuke which might have some insights, at least on how we could hunt this.

[jasagredo]

Closing this one as the io-manager flag fixed it also in CI.